MCP vs Skills: Which Is Right for Your AI Agent and LLMs? Transcript, AI Summary & Key Points

00:00 If you're looking to add capabilities to an LLM, look no further than MCP servers and skills. Because while an LLM is very general purpose and it knows a lot of things, well, I want to show you how these two concepts can help you to add custom and unique data to your LLM, whether it's for all types of use cases, right? This could be for coding assistance, being able to add extra information and data.

00:27 Or if you're building some type of AI agent, right? Or many other use cases. So you're gonna learn whether to choose between MCP or skills or both. Now, let's come back to the LLM, which is essentially like a crystal ball prediction machine that's been trained on all types of information. So books, magazines, even maybe your internet threads might be on one of these LLMs, but it can use this information and pattern recognition to answer tons of different questions, right.

00:56 So maybe one of our questions is, hey, what's the history of Red Hat? Or for us, it could be, how do we inspect a specific database? And so it knows a lot of information, and it might know how to work with your specific database. But getting the right answer means providing the right context to the LLM. So that context might include additional types of information like the way that we want our data formatted.

01:22 Or it could also be the way that our database is configured for our specific team. And maybe we're even going to have some type of tool response that actually went out, got information from that database, and gives it back to the context window of the LLM. So when we think about it, well, this part right here where we're actually just asking a question with a specific role in a task, this is what's known as prompt engineering.

01:48 Just itself, right? We're giving a task to the LLM, but when we add in all of this additional information, this is where it gets really interesting because this is not just prompt engineering, this is what's known as context engineering. So giving the right context to a model to help it to make the right answer, right. So we're giving it all of the information it needs, but the big question here is how do we give the AI model the correct context to make that answer?

02:20 And how do we take this and go build an AI agent from this? Well, those are two great questions. Let's get into it. Let say that an agent needs data that is currently in a customer relationship manager or a CRM. Well, think about the LLM that we're using when we're trying to build an agent. Well instead of giving the LM a copy and paste of this specific CRM or services API and their documentation and a token and saying, hey, please update our customers contact information, make no mistakes, and crossing our fingers.

02:55 Well, MCP or the Model Context Protocol instead standardize how your AI model talks to these different data sources. So what it does is it abstracts your services APIs into a simple LLM ready format as well as for fighting. Authentication, which is really important, for your agent to be able to make the calls to that service. And this provides a uniquely scoped token, read-write access, whatever you need.

03:23 And behind the scenes, the MCP server, when we're working with the agent here, is essentially being added to your IDE, your AI application, and telling the LLM to provide specific JSON request for the information it needs from the service. And that MCP server translates it into post or get requests, so we're then calling that service itself. And MCP is the standardized layer between the LLM and the data sources that you need when you're using an LLM or building an agent, and it's supported by almost all AI tools that

03:52 are out there. The thing is, while MCP solved the problem of how do we give external data to an LLM, there's only one part missing. How do we give that LLM domain knowledge that it might not already have? Because sure with MCP, we can pull the customer records from a CRM, but our sales team wants it done the exact same way every time. And if you know anything about LLMs, you know they're non-deterministic, which makes that really hard.

04:24 So let's say for example, that our sales teams wants the CRM data formatted in a way that has the customer's name, It has the customer's contact information and, very importantly, their favorite type of cookie. So being able to have a repeatable way to say, hey, I want this formatted the same way every time over and over again is the basis of why skills are so important.

04:52 And it works across all major AI tools and models. So think of the task that you use an LLM for, like cleaning up Excel documents. Debugging code, for example, that you always use made to verify or maybe running compliance checks. All of these different prompts and the scripts that you use for them can be packaged into a skill, which is essentially a markdown file with a bit of extra metadata and this is packaged into a folder.

05:23 So this skill, let's talk about it. It's got a formatting of the title, so what the skill is named, the description of when to use it, and under that, the actual prompt itself that will be used and passed to the LLM. And what's really unique about it is the skill can be auto-loaded into the context window of the LLM as needed. So for example, if we had this code debugger skill, we're only going to load it in when asking about code errors.

05:53 And within this folder we can also add different types of resources and different types of scripts. So that if we need additional context or capabilities we can add those to a skill and that's going to be auto loaded in when you're asking about a specific capability that you need from the model. Now when should you pick one or the other? Well for situations where your AI application needs access to real time data but in a controlled and tightly-permissioned way, well, MCP is gonna be the way to go.

06:25 Think of it as an integration between agents and tools, and all you can do is call the tools and resources it provides. So you would use MCP for situations like what VMs am I currently running? Or how's our cluster state? Or what's some information about one of our customers? But the thing is, setting up and configuring an MCP can be quite overkill if you just need to add a reusable and custom capability to your AI.

06:51 And that's where skills shine. Because they're lightweight, they teach your model how to do something. For example, fetch investment data and analyze it, or use scripts or examples that are also included in the skill to perform this task. Both of these can enhance the context window of your LLM, which can help give you the right answer or output you're looking for when building or using AI agents.

07:16 But what's important is that they're both open source, they're both commonly adopted in most AI tools today, and you can start using them today locally on your machine. One thing we didn't talk about when we were comparing both of these is using CLIs for building agents. So, let us know in the comments section below if that's of interest and we'll make a video about it.

07:38 But, until then, be sure to smash that like button if you learned something today and want to hack the algorithm. And let us know if you prefer using MCP servers or skills in the comment section below. We'll see you in the next one.